LayerNorm backward pass produced Inf values input variance below machine epsilon

LayerNorm produced non-finite gradients because the input variance underflowed. The backward pass divides by sqrt(var + eps); in FP16 a near-constant activation drives variance below representable range and the division explodes. This entry explains how to confirm the cause, apply the fix, and separate it from adjacent numerics failures.

The literal signature is "LayerNorm backward pass produced Inf values input variance below machine epsilon". It is a training stability failure associated with mixed precision, gradient scaling, and numerical kernels. The line identifies the failing operation or subsystem, while the surrounding evidence decides whether it is the initiating fault or a downstream symptom.

LayerNorm produced non-finite gradients because the input variance underflowed. The backward pass divides by sqrt(var + eps); in FP16 a near-constant activation drives variance below representable range and the division explodes. The failure becomes visible at this call site because the operation first requires the missing resource, valid state, healthy peer, or correct result. Earlier log lines and a known-good control carry more causal value than the final wrapper exception.

• The workload stops or loses forward progress after emitting "LayerNorm backward pass produced Inf values input variance below machine epsilon".
• A retry on the same configuration reproduces the failure because the causal state has not changed.
• The outer framework exception can hide the rank, node, allocation, or dependency that failed first.
• Increasing timeouts or reducing workload size can suppress the symptom without correcting the cause.

• mixed precision, gradient scaling, and numerical kernels
• production-shaped multi-accelerator workloads
• containerized and bare-metal deployments of the same stack

Apply the following checklist to a small reproduction: each box below is a positive signal that you are looking at this exact failure rather than a sibling in the same taxonomy.

• ✓Find the first occurrence of "LayerNorm backward pass produced Inf values input variance below machine epsilon" and preserve at least 100 lines before it.
• ✓Identify which rank, node, device, or process emitted the first related warning.
• ✓a near-zero variance means the activations entering that layer are nearly constant. That is usually upstream: a dead residual branch, an over-aggressive initialization, or a preceding activation saturating. Inspect the input statistics to the failing layer rather than the layer itself.
• ✓Repeat the same input after the targeted change and require the signature to disappear.
• ✓Resume from last checkpoint before the loss went non-finite only after the control passes.

LayerNorm backward pass produced Inf values input variance below machine epsilon

Timestamps and exact values vary across runs, but the pattern. An info-level start, an early WARN, an ERROR carrying the symptom. Is the actual fingerprint you should alert on. The Denpex platform flags this combination automatically.

raise the layer norm epsilon (1e-5 → 1e-4 for FP16), and run LayerNorm in FP32 even under autocast, it is cheap and is the standard fix. This changes the condition in the causal diagnosis instead of hiding the outer exception. The repeated control proves ownership before recovery from last checkpoint before the loss went non-finite.

# Preserve evidence before restarting
rg -n -i 'error|exception|timeout|failed' <log-file>
nvidia-smi
python -m torch.utils.collect_env

# Find the exact signature in the complete log
rg -n -F -- "LayerNorm backward pass produced Inf values input variance below machine epsilon" <log-file>

Adapt the snippet to your framework. The same pattern holds for PyTorch Lightning, Hugging Face Trainer, DeepSpeed, Megatron-LM, and vLLM training wrappers. Where the wrapper exposes a config flag (for examplelr_scheduler_type in Trainer), prefer the flag over the imperative API to keep the schedule declarative and reproducible.

Select the traceback or open the failed terminal, then run Denpex locally to see the initiating rank, collateral failures, exact fix, and verification command without uploading the log.

• LayerNorm produced non-finite gradients because the input variance underflowed. The backward pass divides by sqrt(var + eps); in FP16 a near-constant activation drives variance below representable range and the division explodes.
• The decisive evidence is the first log line that precedes "LayerNorm backward pass produced Inf values input variance below machine epsilon" and differs from a healthy run.
• A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.